Large language models are increasingly deployed as autonomous agents that communicate, commit, and coordinate in multi-agent systems. Deception in such settings, including promise-breaking, selective information sharing, and exploitation of other agents' interpretive frameworks, introduces deployment risks that isolated-model evaluation cannot detect. Existing evaluations of multi-agent LLM deception are fragmented across subfields that do not share benchmarks or vocabulary, and the resulting measurements rarely compose into a coherent characterization of how frontier models behave. This thesis develops a unified framework for measuring LLM deception in multi-agent settings and populates it with empirical evaluations across three interaction structures. A taxonomy organized along goal-directedness, object, and mechanism dimensions unifies the fragmented literature and reveals systematic benchmark coverage gaps. Three empirical chapters evaluate frontier LLMs in progressively less structured settings: one-shot games with mandated announcements, repeated games with endogenous announcements and heterogeneous model compositions, and a resource-gathering simulation with narrative goals and no announcement protocol. Across these settings, aggregate lying rates obscure qualitatively distinct deceptive profiles. Deception in repeated games with prescribed protocols is predominantly premeditated and takes the form of planned false commitments; under narrative goals and free-form communication, the character of deception varies with goal composition, and the premeditation that occurs takes the form of strategic silence that message-level classification cannot observe. Three candidate monitoring approaches drawn from the existing literature each fail against a specific failure mode. The central claim is that LLM deception in multi-agent settings is not a single phenomenon but a family of structurally distinct failure modes, each shaped by different features of the interaction. Current benchmarks and monitoring approaches systematically underrepresent this variety.

184 pages

Thesis Committee:
Vincent Conitzer (Chair)
Aditi Raghunathan

Jignesh Patel, Interim Head, Computer Science Department
Martial Hebert, Dean, School of Computer Science

Return to: SCS Technical Report Collection
School of Computer Science

This page maintained by reports@cs.cmu.edu